PIVOT: Map-Identity Cross-Fitting for Thermal-to-Visible Localization
Year: 2026
Type: article
Abstract: Reliable thermal-to-visible localization requires verifying whether geometrically consistent matches identify the correct map region. We present Provenance-Isolated Verification of Observation Transport (PIVOT), a predictive verifier that selects among direct, temporally transported, and fused correspondence sources. All occurrences of each persistent map-point identity are assigned to the same fold. Candidate poses are fitted on complementary folds and assessed on held-out observations using a marginalized inlier–outlier likelihood, followed by an all-observation refit of the selected source. On 898 aerial Multi-View Thermal–Visible (MTV) queries, PIVOT achieves 67.82% success at 0.02d/5°, improving direct localization by 5.68 percentage points and unconditional fusion by 3.90 points. It also exceeds the tested transport-aware bundle-adjustment control using the same transported map points. On 300 disjoint Vision for Visibility Dataset (ViViD++) queries, the predeclared PIVOT-Seq obtains 63 successes at 10 m/5°, compared with 61 for direct localization and 62 for fusion. The exploratory core analysis also obtains 63 primary successes, numerically exceeding fusion. The MTV improvement is obtained through per-frame source selection without temporal decoding or scene-specific selector training. The in-sample MTV selector reaches 628 primary successes, so cross-fitting protects against identity reuse rather than maximizing the measured accuracy. The demonstrated gains and the small exploratory external advantage warrant further independent aerial and sensor validation.
Source: Aerospace
Authors: Tonghui Zhang, Peng Han
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7225635933
FWCI:
Cites:
Cited by:
Related to: